Research On Mitosis Detection Based On Breast Pathology Images

Breast cancer is one of the common malignant tumors,and it has been increasing year by year in the female group.The number of mitotic under a microscope cells in a specific area of a breast pathology image is an important indicator of breast cancer classification.At this stage,mitosis counting is performed manually.To identify mitotic cells manually,the operator needs to count under a high-power microscope.This process is not only time-consuming,but also has high professional requirements for the operator.Therefore,it is necessary to design an automatic mitotic detection algorithm based on computer technology as an auxiliary medical method.The current research on mitotic automatic detection algorithms is mainly focused on the feature extraction and classification training stages,which not only takes a lot of time but also has low detection accuracy.Aiming at the different morphology of mitotic cells and the complex background of pathological images,the main work is as follows:Aiming at the problem that the traditional mitosis recognition method directly processes the original RGB image and it is easy to ignore the effective feature information.This paper proposes a mitosis detection method based on multi-channel feature fusion.First,stain normalization method was used for breast cancer pathological images as image pre-processing;Then the image was converted to blue ratio channels to segment the nucleus,and different features of the cells were extracted on multiple color channels for fusion;Finally,support vector machine(SVM)was applied to gets the classification results.The results show that the proposed algorithm is superior to other traditional mitotic recognition algorithms.The fusion of features can improve the detection effect and has higher robustness.Aiming at the problems that traditional mitotic recognition methods need to design complex features and it is difficult to find effective feature expressions,a YOLOv3-based deep learning algorithm is proposed for mitotic detection of breast pathological images.First,manual labeling method was used to complete the target frame labeling for the MICCAI-TUPAC 2016 MITOSIS dataset that only provides the mitotic centroid position;Then uses a feature pyramid network(FPN)to fuse the feature maps at three scales;the feature pyramid network(FPN)was used to fuse the feature maps at three scales;Finally top features are sampled to fuse with the shallow features for detecting mitosis target in three scales respectively.Experimental results show that the proposed algorithm enables the network model to learn more shallow feature information and improve the mitotic detection effect.Its Precision value is 85%more than YOLOV2-416,and the FPS indicator value is 23%more than YOLOV3-416,Indicating that the proposed algorithm achieves higher speed at the same time as higher accuracy.